Feature Overview
Beyond its security model and channel support, Triggerfish provides capabilities that extend your AI agent beyond question-and-answer: scheduled tasks, persistent memory, web access, voice input, and multi-model failover.
Proactive Behavior
Cron and Triggers
Schedule recurring tasks with standard cron expressions and define proactive monitoring behavior through TRIGGER.md. Your agent can deliver morning briefings, check pipelines, monitor for unread messages, and act autonomously on a configurable schedule -- all with classification enforcement and isolated sessions.
Notifications
A notification delivery service that routes messages across all connected channels with priority levels, offline queuing, and deduplication. Replaces ad-hoc notification patterns with a unified abstraction.
Agent Tools
Web Search and Fetch
Search the web and fetch page content. The agent uses web_search to find information and web_fetch to read web pages, with SSRF prevention and policy enforcement on all outbound requests.
Persistent Memory
Cross-session memory with classification gating. The agent saves and recalls facts, preferences, and context across conversations. Memory classification is forced to session taint -- the LLM cannot choose the level.
Image Analysis and Vision
Paste images from your clipboard (Ctrl+V in CLI, browser paste in Tidepool) and analyze image files on disk. Configure a separate vision model to automatically describe images when the primary model does not support vision.
Codebase Exploration
Structured codebase understanding via parallel sub-agents. The explore tool maps directory trees, detects coding patterns, traces imports, and analyzes git history -- all concurrently.
Session Management
Inspect, communicate with, and spawn sessions. The agent can delegate background tasks, send cross-session messages, and reach out across channels -- all under write-down enforcement.
Plan Mode and Task Tracking
Structured planning before implementation (plan mode) and persistent task tracking (todos) across sessions. Plan mode constrains the agent to read-only exploration until the user approves the plan.
Filesystem and Shell
Read, write, search, and execute commands. The foundational tools for file operations, with workspace scoping and command denylist enforcement.
Sub-Agents and LLM Tasks
Delegate work to autonomous sub-agents or run isolated LLM prompts for summarization, classification, and focused reasoning without polluting the main conversation.
Agent Teams
Spawn persistent teams of collaborating agents with specialized roles. A lead coordinates members who communicate autonomously via inter-session messaging. Includes lifecycle monitoring with idle timeouts, lifetime limits, and health checks. Best for complex tasks that benefit from multiple perspectives iterating on each other's work.
Rich Interaction
Voice Pipeline
Full speech support with configurable STT and TTS providers. Use Whisper for local transcription, Deepgram or OpenAI for cloud STT, and ElevenLabs or OpenAI for text-to-speech. Voice input passes through the same classification and policy enforcement as text.
Tide Pool / A2UI
An agent-driven visual workspace where Triggerfish renders interactive content -- dashboards, charts, forms, and code previews. The A2UI (Agent-to-UI) protocol pushes real-time updates from the agent to connected clients.
Multi-Agent and Multi-Model
Multi-Agent Routing
Route different channels, accounts, or contacts to separate isolated agents, each with its own SPINE.md, workspace, skills, and classification ceiling. Your work Slack goes to one agent; your personal WhatsApp goes to another.
LLM Providers and Failover
Connect to Anthropic, OpenAI, Google, local models (Ollama), or OpenRouter. Configure failover chains so your agent automatically falls back to an alternate provider when one is unavailable. Each agent can use a different model.
Rate Limiting
Sliding-window rate limiter that prevents hitting LLM provider API limits. Tracks tokens-per-minute and requests-per-minute, delays calls when capacity is exhausted, and integrates with the failover chain.
Operations
Structured Logging
Unified structured logging with severity levels, file rotation, and dual output to stderr and file. Component-tagged log lines, automatic 1 MB rotation, and a log_read tool for accessing log history.
All features integrate with the core security model. Cron jobs respect
classification ceilings. Voice input carries taint. Tide Pool content passes through the PRE_OUTPUT hook. Multi-agent routing enforces session isolation. No feature bypasses the policy layer. :::